When an Anthropic researcher resigns and uses the words "gambling with our lives," you can dismiss it as one person's nerves, or you can notice who keeps leaving. I notice who keeps leaving.
On 8 September 2026, Jacob Coxon announced he was quitting Anthropic. He'd spent about three years training frontier models, first at OpenAI, then here. His line, as reported by NPR and the Washington Post: the labs "are racing straight to self-improving superintelligence and gambling with our lives." In a farewell note to colleagues he reportedly warned that without more caution and cooperation, superintelligent AI carried "a risk of causing human extinction."
I want to be fair to him, because he was fair. He didn't say Anthropic is evil.
What he actually said
Coxon's read is more careful than the headline. He said he believes Anthropic currently takes safety seriously, but that it's locked in a race, against OpenAI, against Chinese labs, where getting there first is the pressure that eventually bends everything else. "Neither company is acting responsibly," Fortune quoted him, with OpenAI staff who "have not deeply internalized the civilizational stakes" and Anthropic staff who understand the risks but ship anyway because the alternative is losing.
That's the part I'd underline. The problem he describes isn't a villain. It's an incentive. And incentives don't resign.
The pattern nobody wants to name
Here's why one post got my attention. It isn't the first.
In February 2026, Mrinank Sharma, who led Anthropic's Safeguards Research team, resigned publicly with a letter that opened "the world is in peril." Go back to 2024 and OpenAI lost Jan Leike and Ilya Sutskever on the same day, the Superalignment team dissolved behind them, Leike saying safety culture had "taken a backseat to shiny products." Leike then joined Anthropic. Daniel Kokotajlo walked away from roughly $1.7M in equity rather than sign a non-disparagement clause.
Line those up. The people hired specifically to worry about this are the ones with the shortest tenure. When the safety staff are the flight risk and the shipping staff are the retention win, that tells you what the company optimises for, whatever the blog says.
The thing the timing makes hard to ignore
Anthropic's whole founding pitch was that it would be the careful one. Then in February 2026 it quietly rewrote its Responsible Scaling Policy, replacing a categorical commitment to pause if models outran its ability to keep them safe with a conditional one: it would only pause if it both lacked a clear lead over competitors and faced a material catastrophic risk. Read that twice. Competitor behaviour is now an input to whether they slow down for safety.
Meanwhile the revenue chart went vertical. Dario Amodei said the company blew past its "10x per year" plan and hit 80x, a $30B run-rate by April 2026. A company burning at that scale, chasing an IPO, does not get to profitability by being the one that pauses. That's not a conspiracy. It's arithmetic, and the arithmetic and the mission point in different directions now.
Coxon was gentle about it. I don't have to be. When safety is the founding differentiator and the safety people keep quitting while the pause clause gets softened and the valuation climbs, the simplest explanation is that the mission became the marketing and the burn rate became the boss.
I hope he's wrong. I don't think the people still inside get to find out on our behalf.
Top comments (0)